Google DeepMind introduced Gemini Robotics 2.5 on November 10, calling it its most capable vision-language-action model to date. Across more than 200 tasks, DeepMind reports success rates 30 to 100 percent higher than its previous baseline, with tasks finishing 20 to 30 percent faster. The model generalizes across robot bodies, from Apptronik's Apollo 2 humanoid with a five-finger hand to Franka Duo dual-arm systems.

The headline test was not in a lab. With Apptronik and Mercedes-Benz, Apollo 2 humanoids running Gemini Robotics 2.5 performed automotive kitting — sorting parts from bulk bins into trays in the exact order each vehicle needs — on active production lines in Kecskemét, Hungary, and Sindelfingen, Germany, at a reported 96 percent success rate. DeepMind also released ASIMOV-Jailbreak, an open-source benchmark for robot safety prompts.

Treat the 96 percent as a vendor proof-of-concept figure: DeepMind's post does not say how many trials it covers, how success was defined, or how often a human intervened, and there is no independent assessment or Mercedes-side confirmation yet. Gemini Robotics 2.5 is rolling out to robotics partners only — there is no public API, price or general-availability date.